Yep. Capture the distribution you want to sample from by training a DNN on the data and then sample it via generative/diffusion. Basically the DNN is a learnt parametrized distribution and replaces the Markov chain. The adversarial training approach seems to me like a forward backward algorithm over the dataset anyway.